accumulo-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Josh Elser <>
Subject Re: Thrift TTransportException.END_OF_FILE
Date Fri, 12 Jul 2019 17:56:55 GMT
"Commits are held" can be for a couple of different reasons, some from 
within Accumulo and some from outside.

In general, there is an expected ordering of mutations that a 
TabletServer has to apply. A "commit" here is the application of some 
mutations by a TabletServer to the memory map and the WAL.

This could be completely normal and you have some clients which are just 
writing "faster" than your TabletServers can keep up with. This could be 
indicative of slow flushes from memory maps to HDFS. This could be GC 
pressure causing slowness in the TServer.

I'd suggest to take a step back:

* Look at other messages in the DEBUG log for the tabletserver to see if 
you Accumulo is telling you what it's waiting on (before and after you 
see the message about commits being held)
* Check that you're using the Accumulo native memory maps
* Sanity-check performance of HDFS
* Get a thread dump from a TabletServer in this state.

If the problem truly only happens on two servers, it might indicate some 
bad hardware on that device (memory with errors, a disk that flips to r/o).

- Josh

On 7/12/19 10:57 AM, James Srinivasan wrote:
> Hi all,
> We have a Kerberized Accumulo 1.7.0 (HDP3) cluster with 25 tservers.
> Recently, a couple of clients were reporting errrors writing data (fat
> fingered from cluster, apologies for typos):
> org.apache.accumulo.core.client.impl.TabletServerBatchWriter.checkForFailures
> ...
> Caused by: org.apache.accumulo.core.client.impl.TabletServerBatchWriter$MutationWriter.sendMutationsToTabletServer
> Digging into the logs on the problematic tservers, I think the
> following was firing, but don't know why:
> Also, the tserver logs report:
> Internal error processing closeUpdate....TException: Commits are held
> For now, I have stopped the two problematic tservers but any help
> debugging would be much appreciated.
> James

View raw message